Release Notes

Mdapy 1.0.6 (May 6, 2026)

🏆 New Features

  • New mdapy.SQS for Special Quasirandom Structure generation.

  • New mdapy.BondStiffness for fitting longitudinal / transverse bond stiffnesses vs bond length.

  • mdapy.RadialDistributionFunction adds a streaming kernel for arbitrary cutoffs (no Verlet list materialised).

  • mdapy.StructureFactor rewritten in the Faber-Ziman convention of [Erhard2024], with X-ray / neutron / electron totals and the corresponding real-space PDFs.

  • Partial outputs of RadialDistributionFunction / StructureFactor are now dicts keyed by element-pair tuples.

  • New mdapy.unwrap_trajectory() / Trajectory.unwrap().

  • Extended-XYZ I/O recognises GPUMD’s unwrapped_position:R:3 and aliases it to xu/yu/zu.

  • Unified thread-count control via the MDAPY_NUM_THREADS environment variable (set before import mdapy); drives both OpenMP and polars. Defaults to os.cpu_count(). Query via mdapy.get_num_threads(). The public workers / num_threads arguments on mdapy.compress_file and mdapy.render.TachyonRender are removed.

  • New System.cal_chill_plus() (CHILL+ algorithm) for identifying hexagonal ice, cubic ice, gas hydrate and interfacial phases in water simulations.

🐞 Bug Fixes

  • read_dump no longer fails (or silently overwrites) when a dump contains explicit x y z together with xs ys zs and/or xu yu zu. Cartesian columns now take priority and the scaled/unwrapped variants stay under their original names; the fallback promotion only triggers when x y z is absent. Applies to both single-frame mdapy.System and multi-frame mdapy.Trajectory loaders.

Mdapy 1.0.5 (May 1, 2026)

🏆 New Features

  • build_crystal extended to match atomsk --create byte-for-byte: added sc, rocksalt, cscl, zincblende, fluorite, L1_2, perovskite, wurtzite, graphite and lonsdaleite; multi-species via tuple name; Miller orientation for every cubic and Miller-Bravais [hkil] for every hexagonal structure. hcp and graphene now use atomsk’s 2-atom hex primitive (atom counts halve compared with the legacy supercell).

  • New mdapy.orthogonal_cell(system, find_minimal=False) — atomsk’s -orthogonal-cell option.

  • New unified mdapy.Trajectory class for multi-frame XYZ + LAMMPS dump (read + write). Supports fancy indexing with integer arrays and boolean masks (e.g. traj[traj.get_atoms_count() > 100]), optional fast_mode=True for vectorised XYZ reading (~5–7× speedup on regular files), an in-place progress bar (verbose=True, default on), and a vacuum= option on save("out.xyz", vacuum=200.0) that pads every non-periodic axis by vacuum Å and centres the cluster — useful for auto-boxing classical-XYZ training-set frames so downstream MD code sees a well-defined supercell.

🐞 Bug Fixes

  • Neighbor C++ kernel now bounds-checks every write: an over-tight max_neigh raises ValueError instead of silently corrupting memory.

  • System: mutable-default global_info no longer leaks across instances; cal_centro_symmetry_parameter honours the user’s N; cal_structure_entropy keeps both columns when average_rc>0; cal_cluster_analysis raises TypeError on bad rc; write_data no longer mutates self.data; cal_radial_distribution_function no longer overwrites the user’s type column.

  • BuildSystem.from_ovito tolerates sources without a particle_type table.

  • IO writers: write_xyz Properties string assembly; write_dump preserves the element column; cross-device temp files on Windows.

🛠️ Other Improvements

  • Rewrote and optimized the k-nearest-neighbor algorithm; significantly faster than the previous implementation across all tested geometries.

  • API: c_over_a removed from build_crystal / build_hea — pass c directly. System.box is now a property whose setter invalidates neighbor / Voronoi / calculator caches automatically. System.update_data: misspelled reset_calcolator renamed to reset_calculator (old name deprecated).

  • Hardened IO: rewrote LAMMPS data / dump / XYZ / POSCAR / .mp readers to be format-robust (multi-space, CRLF, abc-origin triclinic, Masses element comments, image flags, etc.) and self-consistent on round-trip. Neighbor / NearestNeighbor constructors validate inputs.

  • Internal cleanup: 14 _enlarge_data branches collapsed into one helper; cal_* methods uniformly route through update_data so the DataFrame stays single-chunked; legacy _get_basispos_and_box_cubic removed.

  • Tests: +85 cases (test_system.py, test_build_crystal.py, test_orthogonal_cell.py, test_io_*); IO references match atomsk byte-for-byte. CI is fully self-contained.

  • Trajectory IO speedups: per-frame XYZ parser now picks pl.read_csv (uniform single-space) or numpy split + per-column pl.Series (multi-space); the previous Python dict-of-lists path was dropped.

  • Reworked the entire test suite to be self-contained: reference values for every algorithm are pre-computed and committed as small .npz fixtures under tests/fixtures/. Running the tests no longer requires OVITO, freud, ASE, pymatgen, pynep, or LAMMPS.

  • Added a GitHub Actions workflow that runs the suite on every push and pull request (Windows + Linux + macOS, Python 3.11 / 3.12).

Mdapy 1.0.4 (April 24, 2026)

🏆 New Features

  • Added support for rendering structures using the Tachyon method, accelerated by multi-core CPU and GPU (an order of magnitude faster).

  • Added System.cal_build_bond() to generate bond index arrays directly from neighbor information, supporting a single cutoff as well as type-pair and element-pair cutoffs.

  • Spline now supports three boundary conditions matches scipy.interpolate.CubicSpline.

🐞 Bug Fixes

  • Fixed the EAM potential file reader for LAMMPS eam/alloy format.

  • Fixed EAM interpolation past the tabulated rho range.

🛠️ Other Improvements

  • Added tests/test_spline.py comparing all three boundary conditions against scipy.interpolate.CubicSpline to floating-point precision.

  • Refactor documentation, remove all warnings and errors.

  • Add copyright for all codes.

Mdapy 1.0.3 (March 28, 2026)

🛠️ Other Improvements

  • Added explained_variance_ratio for PCA.

  • For these analysis methods (CNA, PTM, and IDS), simulation boxes with a thickness below 15 Å are automatically treated as small boxes.

  • The optimized elastic constant workflow can be validated using pymatgen.

  • Improved the calculator setup behavior for the System.

  • Fixed a bug in the minimize class.

  • Updated parts of the documentation.

Mdapy 1.0.2 (March 4, 2026)

🛠️ Other Improvements

  • Reduced the size of the published source distribution (sdist).

  • Added support for writing bec, stress, and virial fields in the comment line when using write_xyz.

  • Improved LammpsPotential handling of general simulation boxes; it can now correctly compute stress, forces, and energy for such structures.

  • Added newest version of NEPCPU for fixing a bug in qNEP.

Mdapy 1.0.1 (March 2, 2026)

🏆 New Features

  • Added support for converting VASP OUTCAR files to XYZ format for GPUMD. This supports single-point calculations, structure minimization, and AIMD multi-frame trajectories.

  • Added functionality to identify chemical species based on vdW-radius connectivity between atoms.

🛠️ Other Improvements

  • Fixed a bug in the HEA model builder.

  • Optimized the cfg2xyz conversion tool.

  • Added a neighbor-search benchmark to the documentation.

  • Improved the energy minimization workflow.

  • Optimized stress computation in LammpsPotential and added corresponding tests.

  • Added support for saving BEC information in XYZTraj.

  • Added tests for qNEP.

  • Removed warnings in test_void by introducing an explicit implode cast method.

🚀 Build System

  • Binary wheels are now built separately for each Python version, instead of using the stable ABI (> Python 3.12). This change ensures compatibility with Python 3.13, where the stable ABI is not supported.

Mdapy 1.0.0 (January 3, 2026)

🚀 Summary

This is a milestone update for mdapy, featuring a near-complete rewrite of the core architecture. A primary driver for this transition was the limitations of our previous JIT dependency, Taichi; its development pace constrained our support for newer Python versions.

To ensure long-term sustainability, we have reconstructed mdapy by migrating computationally intensive kernels to C++ using the modern nanobind wrapper. As a result, mdapy now depends solely on NumPy and Polars, making it exceptionally lightweight and compatible with all modern Python environments.

Key changes include:

  • Engine Shift: With the removal of Taichi, mdapy now focuses on high-performance CPU computation.

  • GUI & Tools: Experimental Polyscope support has been removed to focus on core stability. However, a lightweight Jupyter-based GUI remains available as an optional dependency.

  • Modern Build System: We have transitioned from setup.py to pyproject.toml.

  • Reliability: Extensive test suites have been added to ensure the correctness of all algorithms.

This is a brand-new foundation for the project, and we strongly recommend all users to upgrade.

🏆 New features

  • qNEP Integration: Support for evaluating energy, force, virial, charge, and BEC properties.

  • Structural Analysis: Added Static Structure Factors and the Wigner-Seitz method for point defect detection.

  • Mechanical Properties: Support for calculating elastic constants.

  • I/O Enhancements: Added XYZ trajectory loading and introduced the MP file format (Parquet-based), providing high-speed I/O and efficient storage.

  • Minimization: Improved FIRE2 method for energy minimization with cell optimization.

  • GPUMD Ecosystem: A series of new features compatible with the GPUMD economy.

🛠️ Other improvements

  • Compatibility: Full support for Python >= 3.9.

  • Reliability: Significantly expanded test case coverage.

  • Efficiency: Optimized import overhead; import mdapy is now significantly faster.

⚠️ Limitations

  • Documentation: The documentation is currently being updated and is not yet complete. We are actively working on this and welcome any community contributions or feedback.